Research map / Graph learning / Graph self-supervised learning

MAGNET: Manifold-Aware Graph Diffusion Network for Connectome Generation

NeurIPS 2026 · Graph learning › Graph self-supervised learning · approach: Diffusion-based graph generation

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes MAGNET, a graph-transformer diffusion model over a normalized-Cholesky correlation-matrix chart with class-anchored subtype conditioning, an amortized structural bridge, and relevance-preserving corruption, for class-conditional functional connectome generation.

Problem

Generating functional connectomes that are both valid correlation matrices and faithful to class-conditional structure, where coarse binary labels collapse heterogeneous subpopulations and standard validity-preserving generators average across modes.

Approach

Combines a validity-guaranteeing normalized-Cholesky chart with an ROI-token transformer denoiser, and adds three controllers (GMM subtype tokens, a distilled bridge predictor giving structural summaries per step, and condition-adaptive anisotropic noise) to recover class-conditional support under few-step DDIM sampling.

Key results

• Up to 21% relative gain in alpha,beta-F1 over the strongest validity-preserving baseline (21% on OASIS-3, 7% on ADNI), per the paper.
• On zero-shot ADNI subclass evaluation, lambda-MSE reduced by 33-57% relative to baselines (6.08 vs. 9.16 for TriangCFM).
• Amortized bridge gives about 24-33x sampling speedup over teacher-bridge alternatives (Table 5).

Surprising findings

• GDSS-proj, a generic discrete graph diffusion with post-hoc projection, drops to zero alpha,beta-F1 on ABIDE and OASIS-3.

Limitations

• Zero-shot subclass evaluation depends on a classifier-defined subtype space and does not validate biological taxonomy.
• Evaluated only on resting-state fMRI correlation matrices from three cohorts with a 39-region MSDL atlas as main setting.
• Bridge Lipschitz assumption in the theory is a distributional assumption not verified empirically.

Experimental setup

Datasets: ABIDE ADNI OASIS-3

Benchmarks: ABIDE ADNI OASIS-3 AAL116 CC200

Metrics: alpha-precision beta-recall alpha,beta-F1 ROC-AUC F1 lambda-MSE sampling time

Baselines: DiffeoCFM DiffeoGauss TriangCFM TriangDDPM GDSS-proj

Builds on: DiffeoCFM DDIM min-SNR loss weighting classifier-free guidance GDSS

Closest papers by meaning

• Geometry-Guided Generative Representation for Functional Brain Graphs (ICML 2026 · similarity 0.89)
• Riemannian Flow Matching for Brain Connectivity Matrices via Pullback Geometry (NeurIPS 2025 · similarity 0.88)
• BrainFlow: A Holistic Pathway of Dynamic Neural System on Manifold (NeurIPS 2025 · similarity 0.88)
• Exploring the Enigma of Neural Dynamics Through A Scattering-Transform Mixer Landscape for Riemannian Manifold (ICML 2024 · similarity 0.88)
• Long-range Brain Graph Transformer (NeurIPS 2024 · similarity 0.86)
• Patch2Self2: Self-supervised Denoising on Coresets via Matrix Sketching (CVPR 2024 · similarity 0.85)